What Counts as an AEDT? Scoping Local Law 144 Across Your Hiring Stack

Summary: Local Law 144's obligations attach to automated employment decision tools. Employers routinely scope this too narrowly, checking only their applicant tracking system and concluding they are out of scope. The definition in DCWP's final rule is broader than that, and the features that trigger it are usually embedded inside platforms bought for other reasons.

The definition

Under the DCWP final rule at 6 RCNY sections 5-300 to 5-304, adopted April 2023, an AEDT is a computational process derived from machine learning, statistical modelling, data analytics, or artificial intelligence, which issues a simplified output such as a score, classification, or recommendation, used to substantially assist or replace discretionary decision-making for hiring or promotion.

Four elements, all of which must be present.

A computational process from one of four named techniques. Note that statistical modelling and data analytics are in the list alongside machine learning and AI. A tool does not need to be an AI product in the marketing sense. A regression-based scoring model built in 2015 qualifies.

A simplified output. A score, a classification, a ranking, a recommendation, a pass or fail. Something that compresses a candidate into a comparable unit.

Substantially assists or replaces discretion. This is the contested element, and where most scoping arguments happen.

Used for hiring or promotion. Employment decisions, not workforce analytics generally.

Why "substantially assist" is the hard part

The final rule narrowed AEDT scope relative to the original 2022 proposed rule, and employers have leaned on that narrowing more than it supports.

A tool that produces a ranked candidate list which a recruiter then works down from the top is not obviously outside the definition simply because a human clicks the buttons. The question is whether the output substantially assists the discretionary decision. If the ranking determines review order, and review order determines who gets seen in a high-volume funnel, the ranking is doing decision work.

The safer framing for scoping: ask what would change if the tool were switched off tomorrow. If the answer is "we would look at a different set of candidates," the tool is assisting the decision.

Where scope hides in a typical stack

In practice, the flagship applicant tracking system is often not the exposure. These are the recurring surprises.

Match and fit scores inside sourcing platforms. Many sourcing tools surface a percentage match against a job description. If recruiters filter or prioritise on it, that is a simplified output assisting a screening decision.

Assessment providers. Cognitive, behavioural, and situational judgement assessments that return a score or band. These are frequently procured by a different team than the ATS and are missed in inventories built from the recruiting stack alone.

Resume parsers that rank rather than sort. Parsing is not scoring. Ranking is. The distinction is worth checking with the vendor rather than assuming.

Video interview platforms with scoring features. Even where the scoring is optional, the question is whether it is switched on.

Knockout question logic. Rule-based screening built by a human is generally not derived from machine learning, statistical modelling, or data analytics, so simple knockout questions usually fall outside. But automated scoring layered on top of those responses may not.

Internal mobility and promotion tools. Local Law 144 covers promotion, not just hiring. Talent marketplace and succession products are commonly omitted from scoping exercises focused on external recruiting.

Building a defensible inventory

The output you want is not a list of AI tools. It is a list of every point in the funnel where a candidate is scored, ranked, classified, or filtered by something other than a person reading their application.

1. Map the funnel end to end, from sourcing through offer, including internal moves.
2. At each stage, ask what narrows the pool and what mechanism does the narrowing.
3. For each mechanism, record the vendor, the technique, the output type, and whether the output is used. A feature that exists but is disabled is worth documenting as disabled.
4. Ask each vendor directly whether they consider the feature an AEDT under the DCWP rule, and keep the answer. A vendor position is not a legal defence, but a documented inquiry is better than silence.
5. Record scoping decisions with reasoning, including the tools you concluded were out of scope. Under a proactive enforcement posture, the ability to show a considered scoping exercise is worth more than a clean list with no working shown.

The remote hiring trap

Scope under Local Law 144 is set by candidate location at the time of screening. For an employer running remote searches, the practical consequence is that the tool inventory question is not "do we hire in New York" but "could a New York resident enter this funnel."

For most companies posting remote US roles, the answer is yes, and the scoping exercise applies to the whole stack rather than to a New York-specific process.

Frequently asked questions

Is a keyword search an AEDT? Generally no. A boolean search executed by a recruiter is not a computational process issuing a simplified output derived from the named techniques. A relevance-ranked result set from a machine learning model is a different question.

What if the tool only shortlists and a human makes the final decision? Human final decisions do not by themselves remove a tool from scope. The rule covers tools that substantially assist discretionary decision-making, not only tools that replace it.

Does the law apply to tools used only for internal promotion? Yes. Promotion is expressly covered.

Our vendor says their tool is not an AEDT. Is that sufficient? It is evidence, not a defence. The obligation sits with the employer, and DCWP's determination of whether a tool is an AEDT is not bound by the vendor's view. Keep the vendor's position and your own reasoning.

How often should the inventory be refreshed? At minimum annually, aligned to the bias audit cycle, and whenever a new tool or feature enters the funnel. Vendors add scoring features to existing products, which means a tool scoped out last year may be in scope now without any procurement event to trigger a review.